Method for local quality assessment of cryo-em 3d density maps based on deep learning

Abstract

A three-dimensional density map local quality evaluation method based on deep learning, by constructing a DeepQs deep learning model based on VisionTransformer as the backbone network, and taking the local density map of the cryo-electron microscopy three-dimensional density map obtained by resampling as the input, and taking the corresponding atomic structure and the matching resampled three-dimensional density map as the label for training, in the online stage, after resampling the three-dimensional density map of the biological macromolecule model reconstructed by the cryo-electron microscope, the region of interest is obtained through the mask and input into the trained DeepQs deep learning model for local quality evaluation and quality evaluation map is obtained. The present application combines the use of the decomposability score (Q-score) of atoms in the cryo-electron microscopy three-dimensional density map, which can score each local region of the three-dimensional density map and obtain a quality evaluation map.

Description

Technical Field

[0001] This invention relates to a technique in the field of image processing, specifically a local quality assessment method based on a three-dimensional density map obtained by cryo-electron microscopy using deep learning. Background Technology

[0002] Currently, reconstruction techniques for cryo-electron microscopy images primarily use resolution as a numerical metric for evaluation. These include the Fourier Shell Correlation (FSC) method and the ResMap method. The former uses two 3D density maps constructed from half of the cryo-electron microscopy image, then calculates the correlation between the two reconstructed datasets in Fourier space to obtain the FSC curve. A specific threshold is selected, and the value below that threshold represents the global resolution. The ResMap method detects different sinusoidal wavelengths at various points in the 3D density map; the wavelength with the lowest value among all wavelengths and higher than the noise signal is defined as the local resolution at that point. However, the existing FSC threshold selection lacks intuitive meaning, ResMap cannot handle sharpened and filtered 3D density maps, and finally, the evaluation criteria of the aforementioned existing methods are not directly related to the desired reconstructed atomic structure model. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a deep learning-based method for local quality assessment of 3D density maps. By using deep learning, a model is trained that can obtain approximate assessments with only a 3D density map as input. By combining the decomposability score (Q-score) of atoms in the cryo-electron microscopy 3D density map, each local region of the 3D density map can be scored, and a quality assessment map can be obtained.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a deep learning-based method for local quality assessment of 3D density maps. A DeepQs deep learning model with VisionTransformer as its backbone network is constructed. The model is trained using a local density map of a cryo-electron microscopy (CEM) 3D density map obtained through resampling as input, and the Q-score between the corresponding atomic structure and the matching resampling 3D density map as a label. In the online stage, after resampling the 3D density map of the biomacromolecule model reconstructed from the cryo-electron microscopy image, the region of interest is obtained through a mask and input into the trained DeepQs deep learning model for local quality assessment, resulting in a quality assessment map.

[0006] The aforementioned cryo-electron microscopy three-dimensional density map refers to a three-dimensional matrix of size w×h×d, where each element is a voxel, and the coordinates (x, y, z) of each voxel are its relative coordinates in space. The physical meaning of the value of each voxel is the electron density at its coordinate.

[0007] The resampling refers to adjusting the minimum spacing between voxels in the three-dimensional density map of biological macromolecules to 1.0 Å using software to ensure that the density map input to the model is physically consistent.

[0008] The Q-score refers to the correlation between the mapping value of each point around an atom in a registered atomic structure on a 3D density map and a reference Gaussian function. Specifically, for any atom in the map, the surrounding density distribution conforms to... is a Gaussian distribution, where: , , M is a three-dimensional density map. The average density of the 3D density map; in the input (0-2 Within the range of step distance Uniform sampling yields a vector composed of the output values. The Euclidean distance from the center coordinates of this atom within the three-dimensional density map. The positional interpolation sampling density values ​​are used to obtain the vector. ; calculate the vectors respectively sum vector mean and The Q-score of this atom was obtained as follows: .

[0009] The DeepQs deep learning model includes a 3D matrix dimensionality reduction module, a Transformer encoder module, and an output module. The 3D matrix dimensionality reduction module, which takes a 3D matrix as input to adapt to the density map, first divides the input 3D matrix into multiple 3×3×3 3D matrix blocks. Using 3D convolution technology, combined with a 3×3×3 parameter matrix W, the two are convolved, mapping each 3D block to a 1D space. Based on the number of parameter matrices W used (n=128), the 3×3×3 matrix blocks are finally mapped to 1×128 vectors, resulting in 27 blocks and a data dimension of 27×128. The Transformer encoder module, based on a multi-head self-attention mechanism combined with multilayer perceptron, consists of blocks formed by residual connections. By acquiring the connections between the various matrix blocks, it helps the network learn. The output unit takes the output of the Transformer encoder module, passes it through a multilayer perceptron, and then uses the tanh function to obtain the output result.

[0010] The local density map refers to the calculation of the quality assessment score of each sampling point relative to the three-dimensional density map of the entire biomolecule. It does not require global information, but only needs to crop the surrounding area, which is the local density map.

[0011] The sampling points are atomic coordinates within the registered atomic structure model, and the 3D density map uses the same coordinate system as the atomic structure model. A 9×9×9 cube region is expanded outwards from the sampling point to form the local density map. If the expanded region exceeds the specified range, zeros are padded to create a complete 9×9×9 cube.

[0012] The mask refers to a three-dimensional matrix with the same size as the input three-dimensional density map, consisting only of 0s and 1s. The mask separates background noise and negative value regions from the main macromolecule region of interest. The mask is constructed by setting a threshold value, setting regions above the threshold to 1 and regions below the threshold to 0.

[0013] The quality assessment score refers to the score obtained by summing and averaging all DeepQs scores. The score obtained is the overall quality assessment of the entire 3D density map. The closer the score is to 1, the higher the quality score of the 3D reconstruction model in the scoring rule of this invention, and vice versa.

[0014] Technical effect

[0015] This invention applies atomic structure evaluation metrics to the quality assessment of cryo-electron microscopy (cryo-EM) 3D density maps, combined with deep learning. Compared to existing technologies, it can predict the fit between cryo-EM density maps and atomic models without requiring input of atomic structures, thus overcoming some shortcomings of existing techniques. For example, this method does not require inputting two half-dataset reconstructed models as in the FSC method, nor does it require calculating noise distribution; it directly inputs the reconstructed 3D density map and mask threshold, making it simple and fast to use. Finally, compared to other local resolution methods, this invention does not suffer from the problem of being unable to evaluate the sharpened 3D density map. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention;

[0017] Figure 2 This is a schematic diagram of the atomic structure model of a biological macromolecule in an example.

[0018] Figure 3 This is a schematic diagram of a three-dimensional density map obtained by cryo-electron microscopy in an example.

[0019] Figure 4This is a schematic diagram of the quality assessment output from the DeepQs model in an example.

[0020] Figure 5 The following is a curve showing the fitting relationship between the average DeepQs output and the global resolution of multiple cryo-electron microscopy 3D density maps in the example.

[0021] Figure 6 This is a diagram of the DeepQs neural network structure. Detailed Implementation

[0022] like Figure 1 As shown in this embodiment, a local quality assessment method based on a deep learning cryo-electron microscopy three-dimensional density map is used, including:

[0023] The first step involved uniformly selecting three-dimensional density maps of biological macromolecules and matching atomic structure models from a resolution range of 1 Å to 7 Å according to the sampling requirements of the dataset. A total of 495 pairs of data were selected as the basic dataset.

[0024] like Figure 2 and Figure 3 The dataset shows cryo-electron microscopy (cryo-EM) 3D density maps of biological macromolecules and their corresponding matching atomic model structures. There are a total of 495 pairs of cryo-EM 3D density maps of biological macromolecules with their corresponding matching atomic model structures.

[0025] The second step involves using software to calculate the Q-score of each atom in the atomic structure model for each pair of biomacromolecules and the matched atomic structure model.

[0026] The third step involves cropping a 9×9×9 cubic local density map from the corresponding 3D density map of the biomacromolecule based on the coordinates of each atom in the atomic structure model. This local density map is then used as input, and the Q-scores of the selected atoms are used as labels. These are then fed into the DeepQs model for training. The neural network diagram is shown below. Figure 6 As shown, approximately 400,000 local density maps were ultimately sampled as the training set.

[0027] The fourth step is to perform a quality assessment on the novel cryo-electron microscopy 3D density map. In this embodiment, density map numbered 26981 in the EMDB is used for testing. The macromolecule is as follows... Figure 3 As shown, the 3D density map was resampled using Chimera software to obtain a 3D density map with a grid spacing of 1.0 Å.

[0028] The fifth step involves processing the resampled cryo-electron microscopy 3D density map using the recommended thresholds provided by the EMDB website to obtain a mask.

[0029] Step 6: For the resampled 3D density map, using the mask obtained in step 5, obtain the coordinates of all values ​​of 1 in all masks. Using these coordinates as the center, crop out a 9×9×9 local density map from the resampled 3D density map, input it into the trained DeepQs model, and output the quality evaluation of each local density map. Overlay all DeepQs local quality score evaluations onto the corresponding coordinates of the original density map to obtain the quality evaluation map of the 3D density map.

[0030] Through specific practical experiments, under the condition of using a single NVIDIA GeForce GTX 3090Ti GPU, the following were conducted: Figure 3 The EMDB-26981 was used for quality assessment, with an input 3D density map of size 123 × 89 × 101. A mask was created with a threshold of 1.5, selecting 32251 voxels in the 3D density map. Sampling was performed on these voxels, resulting in 32251 corresponding 9×9×9 cubes representing local density distributions. Data was input into the DeepQs model in batches of 256, obtaining the DeepQs quality assessment score for each sampling point. The entire process took 22 seconds, and the calculated quality assessment map is shown below. Figure 4 As shown.

[0031] To compare the correlation between the quality assessment of this invention and global resolution, and for 40 cryo-electron microscopy 3D density maps with different global resolutions, the same steps 1-6 were performed as described above, using the same trained DeepQs model to obtain their respective local quality assessments. The local quality assessments were then summed and averaged to obtain the relationship between the local quality assessments and the global resolution, as shown in the graph below. Figure 5 As shown.

[0032] To demonstrate the effect of this method on the evaluation of the sharpened models, the quality of multiple models before and after sharpening was evaluated. The results are shown in Table 1. If the sharpening results in an increase in the quality score of each model as expected, then sharpening the model will not produce abnormal output.

[0033] Table 1: Average output of DeepQs quality assessment before and after sharpening

[0034] Cryo-electron microscopy three-dimensional density map code Average output (before sharpening) Average output (after sharpening) EMDB-3407 0.387 0.550 EMDB-27802 0.531 0.584 EMDB-33259 0.612 0.635

[0035] Compared to existing techniques, this method uses a novel dataset and labels for training. The training labels are Q-scores. Calculating the Q-score requires atomic models and 3D density maps as inputs. Through deep learning, this method eliminates the need for atomic coordinate models while achieving approximate evaluation results. Secondly, besides serving as a new quality evaluation standard for cryo-electron microscopy 3D density maps, this method also leverages the advantage of Q-scores in assessing the fit between atomic models and 3D density maps. Furthermore, this method can detect differences in quality between cryo-electron microscopy 3D density maps before and after sharpening. Finally, this quality evaluation method, based on deep learning, only requires the 3D density map and mask threshold as inputs, making it simpler and faster to use.

[0036] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for local quality assessment of 3D density maps based on deep learning, characterized in that, By constructing a DeepQs deep learning model based on Vision Transformer as the backbone network, and taking the local density map of the cryo-electron microscopy 3D density map obtained by resampling as the input, and using the Q-score between the corresponding atomic structure and the matching resampled 3D density map as the label for training, after resampling the 3D density map of the biomacromolecule model reconstructed by the cryo-electron microscopy to be tested in the online stage, the region of interest is obtained through a mask and then input into the trained DeepQs deep learning model for local quality evaluation and to obtain the quality evaluation map. The Q-score refers to the correlation between the mapping value of each point around an atom in a registered atomic structure on a 3D density map and a reference Gaussian function. Specifically, for any atom in the map, the surrounding density distribution conforms to... is a Gaussian distribution, where: , , M is a three-dimensional density map. The average density of the 3D density map; in the input (0-2 Within the range of step distance Uniform sampling yields a vector composed of the output values. The Euclidean distance from the center coordinates of this atom within the three-dimensional density map. The positional interpolation sampling density values ​​are used to obtain the vector. ; calculate the vectors respectively sum vector mean and The Q-score of this atom was obtained as follows: ; The DeepQs deep learning model includes a 3D matrix dimensionality reduction module, a Transformer encoder module, and an output module. The 3D matrix dimensionality reduction module, which takes a 3D matrix as input to adapt to the density map, first divides the input 3D matrix into multiple 3×3×3 3D matrix blocks. Using 3D convolution technology, it performs convolution with a 3×3×3 parameter matrix W, mapping each 3D block to a 1D space. Based on the number of parameter matrices W (n=128), the 3×3×3 matrix blocks are finally mapped to 1×128 vectors, resulting in 27 blocks and a data dimension of 27×128. The Transformer encoder module, based on a multi-head self-attention mechanism combined with multilayer perceptron, consists of blocks formed by residual connections. By acquiring the connections between the various matrix blocks, it helps the network learn. The output unit takes the output of the Transformer encoder module, passes it through a multilayer perceptron, and then uses the tanh function to obtain the output result.

2. The method for local quality assessment of 3D density maps based on deep learning according to claim 1, characterized in that, The aforementioned cryo-electron microscopy three-dimensional density map refers to a three-dimensional matrix of size w×h×d, where each element is a voxel, the coordinates (x, y, z) of each voxel are its relative coordinates in space, and the physical meaning of the value of each voxel is the electron density at its coordinate. The resampling refers to adjusting the minimum spacing between voxels in the three-dimensional density map of biological macromolecules to 1.0 Å using software to ensure that the density map input to the model is physically consistent.

3. The method for local quality assessment of 3D density maps based on deep learning according to claim 1, characterized in that, The local density map refers to the calculation of the quality assessment score of each sampling point relative to the three-dimensional density map of the entire biomolecule. It does not require global information, but only needs to crop the surrounding area, which is the local density map.

4. The method for local quality assessment of 3D density maps based on deep learning according to claim 3, characterized in that, The sampling points are atomic coordinates in the registered atomic structure model, and the three-dimensional density map uses the same coordinate system as the atomic structure model. The sampling points are centered on the sampling points and expand outwards to form a 9×9×9 cube region as the local density map of the sampled area. If the expanded area exceeds the range, zeros are added to complete the 9×9×9 cube.

5. The method for local quality assessment of 3D density maps based on deep learning according to claim 1, characterized in that, The mask refers to a three-dimensional matrix with the same size as the input three-dimensional density map, consisting only of 0s and 1s. The mask can separate background noise and negative value regions from the main macromolecule region of interest. The construction of the mask is determined by setting a threshold, setting the value of regions above the threshold to 1 and the value of regions below the threshold to 0.

6. The method for local quality assessment of 3D density maps based on deep learning according to claim 3, characterized in that, The quality assessment score refers to the score obtained by summing and averaging all DeepQs scores to determine the overall quality assessment of the entire 3D density map. The score is the decomposability score of the voxels with a mask value of 1 in the original 3D density map after DeepQs output, which is in the range of -1 to 1.

Images